mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-11-17 04:11:35 +00:00
Snapshot history with copy.deepcopy in the canonicalization test (a shallow copy shares the nested tool_call dicts, so an in-place mutation would pass), and add a list-content row with a lone surrogate to the surrogate test, asserting the whole stored history is unchanged. The surrogate sanitizer is in-place and list parts are nested dicts shared with history, so this row is what pins the _clone_message_for_send step: deleting it turns the test red.
105 lines
5.1 KiB
Python
105 lines
5.1 KiB
Python
"""The max-iterations summary call treats ``reasoning_details`` per api_mode: the
|
|
anthropic_messages converter rebuilds signed thinking blocks from it, so the summary messages
|
|
must keep it; a strict chat-completions route drops it on the wire via the same kwargs
|
|
builder the main loop uses (hermes-agent#70233)."""
|
|
|
|
import copy
|
|
import json
|
|
|
|
import pytest
|
|
|
|
from agent.chat_completion_helpers import _build_api_kwargs_for_mode, _iteration_summary_api_messages
|
|
from run_agent import AIAgent
|
|
|
|
_HISTORY = [
|
|
{"role": "user", "content": "q"},
|
|
{"role": "assistant", "content": "", "tool_calls": [{"id": "t1", "type": "function", "function": {"name": "f", "arguments": "{}"}}],
|
|
"reasoning_details": [{"type": "thinking", "thinking": "x", "signature": "SIG"}]},
|
|
{"role": "tool", "tool_call_id": "t1", "content": "r"},
|
|
]
|
|
|
|
|
|
@pytest.fixture
|
|
def make_agent(tmp_path, monkeypatch):
|
|
monkeypatch.setenv("HERMES_HOME", str(tmp_path))
|
|
|
|
def _make(base_url, provider):
|
|
agent = AIAgent(api_key="k", base_url=base_url, provider=provider, model="m", quiet_mode=True,
|
|
skip_context_files=True, skip_memory=True)
|
|
agent._cached_system_prompt = "SYS"
|
|
return agent
|
|
return _make
|
|
|
|
|
|
def test_anthropic_summary_messages_keep_reasoning_details(make_agent):
|
|
agent = make_agent("https://api.anthropic.com", "anthropic")
|
|
assert agent.api_mode == "anthropic_messages"
|
|
out = _iteration_summary_api_messages(agent, [dict(m) for m in _HISTORY])
|
|
assistant = next(m for m in out if m.get("role") == "assistant")
|
|
assert assistant["reasoning_details"] == _HISTORY[1]["reasoning_details"]
|
|
|
|
|
|
def test_strict_chat_route_summary_wire_drops_reasoning_details(make_agent):
|
|
agent = make_agent("https://api.groq.com/openai/v1", "custom")
|
|
assert agent.api_mode == "chat_completions"
|
|
api_messages = _iteration_summary_api_messages(agent, [dict(m) for m in _HISTORY])
|
|
kwargs = _build_api_kwargs_for_mode(agent, api_messages)
|
|
assert all("reasoning_details" not in m for m in kwargs["messages"])
|
|
|
|
|
|
class TestSummaryPrefixParity:
|
|
"""The summary request must close with the same normalization the main send
|
|
path applies (assemble_api_request): canonical tool-call argument JSON and
|
|
stripped string content — otherwise the summary's prefix diverges from every
|
|
prior request and prefix-caching providers re-read the whole conversation
|
|
(hermes-agent#123002)."""
|
|
|
|
def test_canonicalizes_tool_call_arguments_and_strips_content(self, make_agent):
|
|
agent = make_agent("https://api.groq.com/openai/v1", "custom")
|
|
history = [
|
|
{"role": "user", "content": " q "},
|
|
{"role": "assistant", "content": " thinking out loud ", "tool_calls": [
|
|
{"id": "t1", "type": "function",
|
|
"function": {"name": "f", "arguments": '{"b": 1, "a": 2}'}},
|
|
]},
|
|
{"role": "tool", "tool_call_id": "t1", "content": " r "},
|
|
]
|
|
snapshot = copy.deepcopy(history)
|
|
out = _iteration_summary_api_messages(agent, history)
|
|
|
|
assistant = next(m for m in out if m.get("role") == "assistant")
|
|
args = assistant["tool_calls"][0]["function"]["arguments"]
|
|
assert args == json.dumps({"a": 2, "b": 1}, separators=(",", ":"), sort_keys=True)
|
|
|
|
# Same whitespace contract as the main path: string content stripped on the
|
|
# wire copy, history untouched.
|
|
by_role = {m["role"]: m for m in out if m.get("role") in {"user", "tool"}}
|
|
assert by_role["user"]["content"] == "q"
|
|
assert by_role["tool"]["content"] == "r"
|
|
assert history == snapshot # send-path rewrites never reach the transcript
|
|
|
|
def test_strips_lone_surrogates_like_the_send_path(self, make_agent):
|
|
agent = make_agent("https://api.groq.com/openai/v1", "custom")
|
|
history = [
|
|
{"role": "user", "content": "clip \ud800 paste"},
|
|
{"role": "assistant", "content": "ok \ud83d", "tool_calls": [
|
|
{"id": "t1", "type": "function",
|
|
"function": {"name": "f", "arguments": '{"k": "v\ud800"}'}},
|
|
]},
|
|
{"role": "tool", "tool_call_id": "t1", "content": "r"},
|
|
{"role": "user", "content": [{"type": "text", "text": "part \ud800"}]},
|
|
]
|
|
snapshot = copy.deepcopy(history)
|
|
out = _iteration_summary_api_messages(agent, history)
|
|
# Main path's third pass rewrites lone surrogates to U+FFFD; anything else diverges the
|
|
# prefix or makes the SDK's ensure_ascii=False utf-8 encode raise.
|
|
assert next(m for m in out if m["role"] == "user")["content"] == "clip \ufffd paste"
|
|
assistant = next(m for m in out if m["role"] == "assistant")
|
|
assert assistant["content"] == "ok \ufffd"
|
|
# Canonicalization runs first, so argument surrogates are already ASCII \udXXX escapes.
|
|
assert assistant["tool_calls"][0]["function"]["arguments"] == '{"k":"v\\ud800"}'
|
|
assert out[-1]["content"][0]["text"] == "part \ufffd"
|
|
json.dumps(out, ensure_ascii=False).encode("utf-8")
|
|
# The sanitizer is in-place and list-content parts are shared nested dicts:
|
|
# history must keep its stored bytes.
|
|
assert history == snapshot
|